Systems of equations

What if we have more than one variable that is evolving through time and space?

• populations of two species that are changing through time, and interact with each other

• pollutant concentrations in a diffusion-advection model where the two pollutants also interact with each other

• growth of above-ground and below-ground carbon and nitrogen in a forest

• household consumption and savings through time (when consumption depends on savings and vice versa)

These require several differential equations that have to be solved simultaneously

Systems of equations

Systems of equations

• can be ordinary (differentiating with respect to one variable (time or space))

• partial if differentiating with respect to multiple variables (x,y,z, time)

We can estimate trajectories of systems of equations in much the same way that we used numerical integration for our ODE’s

Here again methods (especially for complicated parital derivative system of equations ) can be complicated

Call your engineering/math when you run into issues!

Dynamics of two (or more) variables

• two variable dynamic models that have feedbacks between variables can create cyclical dynamics (and more complex )

• Two ways to look at results

  • time series of each state variable

  • how state variables interact with each other

Predator-Prey Models

Predator-Prey models

A simple approach that assumes prey grow exponentially, with a fixed intrinsic growth rate

• a fixed mortality rate of predators
• a fixed rate of consumption/predation rate of prey by predators
• a fixed conversion rate (ingestion rate) that determines how many “new” predators you get with predation
• no environmental effects (e.g no carrying capacity)

Predator-Prey Model

Analogs

As with diffusion, the basic form/ideas in this model can be applied elsewhere

• economics

• infectious disease spread

• combustion

Differential equations for a Predator-Prey Model

• Prey

\(\frac{\partial prey}{\partial t} = r_{prey} * prey - \alpha * prey * pred\)

• Predator

\(\frac{\partial pred}{\partial t} = eff * \alpha * pred * prey - mort * pred\)

Predator Prey - Implementation in R

• Ordinary Differential Equation with two dependent variables

• Still use ODE solve in R

• Still code the derivative as a function

• Use lists or vectors to bring in initial conditions for all dependent variables; (similar how we bring in multiple parameters to derivative definition function)

• use with can help make it easier to code the use of parameters within the derivative definition function (see example below)

• use lists to output derivatives for all dependent variable

Example implementation

source("../R/lotvmod.R")
lotvmod

## function (t, pop, pars) 
## {
##     with(as.list(c(pars, pop)), {
##         dprey = rprey * prey - alpha * prey * pred
##         dpred = eff * alpha * prey * pred - pmort * pred
##         return(list(c(dprey, dpred)))
##     })
## }

# note the use of with
# initial conditions
currpop = c(prey=10, pred=1)

# time points to see results
days = seq(from=1, to=100, by=1)

# set parameters
pars = c(rprey=0.5, alpha=0.3, eff=0.2,pmort=0.2, K=100)

# run the model
res = ode(func=lotvmod, y=currpop, times=days, parms=pars)

Visualizing results

• two variable dynamic models that have feedbacks between variables can create cyclical dynamics (and more complex )

• Two ways to look at results

  • time series of each state variable (pred and prey)

  • how state variables interact with each other

Relationship between populations

# graph the results
head(res)

##      time      prey     pred
## [1,]    1 10.000000 1.000000
## [2,]    2 11.320956 1.564997
## [3,]    3 10.244569 2.482924
## [4,]    4  6.914805 3.420960
## [5,]    5  3.777091 3.836373
## [6,]    6  1.987468 3.710744

# rearrange for easy plotting
resl = as.data.frame(res) %>% pivot_longer(-time, names_to="animal", values_to="pop")
p1=ggplot(resl, aes(time, pop, col=animal))+geom_line()

p1

p2=ggplot(as.data.frame(res), aes(pred, prey))+geom_point()+labs(y="Prey",x="Predators")
p2

# To make this easier to understand - maybe
p2b=ggplot(as.data.frame(res), aes(pred, prey, col=time))+geom_point()+labs(y="Prey",x="Predators")
p2b

ggarrange(p1,p2b)

#Try other parameters - try to bring relative size of predictors (versus prey) higher

Other illustations

• Prey

\(\frac{\partial prey}{\partial t} = r_{prey} * prey - \alpha * prey * pred\)

• Predator

\(\frac{\partial pred}{\partial t} = eff * \alpha * pred * prey - mort * pred\)

Predator and Prey with Carrying Capacity?

How would you code that?

With Carrying Capacity

• Prey

\(\frac{\partial prey}{\partial t} = r_{prey} * (1-\frac{prey}{K})*prey - \alpha * prey * pred\)

• Predator

\(\frac{\partial pred}{\partial t} = eff * \alpha * pred * prey - mort * pred\)

Implementation

source("../R/lotvmodK.R")
lotvmodK

## function (t, pop, pars) 
## {
##     with(as.list(c(pars, pop)), {
##         dprey = rprey * (1 - prey/K) * prey - alpha * prey * 
##             pred
##         dpred = eff * alpha * prey * pred - pmort * pred
##         return(list(c(dprey, dpred)))
##     })
## }

# initial conditions
currpop=c(prey=1, pred=1)

# set parameter list
pars = c(rprey=0.1, alpha=0.6, eff=0.8,pmort=0.4, K=20)

# times when you want to evaluate
days = seq(from=1,to=500)

# run our differential equation solver
res = ode(func=lotvmodK, y=currpop, times=days, parms=pars)

# rearrange for plotting
resl = as.data.frame(res) %>% pivot_longer(-time, names_to="species", values_to="pop")

# graph both populations over time
p1=ggplot(resl, aes(time, pop, col=species))+geom_line()
p1

# also look at relationships between preditor and prey population and use color for time 
# I will remove the legend here to make it easier to see 
p2 = ggplot(as.data.frame(res), aes(pred, prey, col=(round(time/10))))+geom_point()+theme(legend.position = "none")
p2

p2 = ggplot(as.data.frame(res), aes(pred, prey, col=as.factor(round(time/10))))+geom_point()+theme(legend.position = "none")
p2

ggarrange(p1,p2)

# try with different parameter sets, can you create one where populations are stable - less cycling?

Competition

Species 1 (or Company 1)

\(\frac{\partial s_1}{\partial t} = r_{1} * s_1 * (1-(\frac{s_1+\alpha_{12} * s_2}{K_1}))\)

Species 2 (or Company 2)

\(\frac{\partial s_2}{\partial t} = r_{2} * s_2 * (1-(\frac{s_2+\alpha_{21} * s_1}{K_2}))\)

How might you explain what this is doing?

Competition

Species 1 (or Company 1)

\(\frac{\partial s_1}{\partial t} = r_{1} * s_1 * (1-(\frac{s_1+\alpha_{12} * s_2}{K_1}))\)

Species 2 (or Company 2)

\(\frac{\partial s_2}{\partial t} = r_{2} * s_2 * (1-(\frac{s_2+\alpha_{21} * s_1}{K_2}))\)

• \(s_1, s_2\) are species populations
• \(r_1, r_2\) are growth rates of each species
• \(K_1, K_2\) are carrying capacities; could be the same for both species but maybe not?
• \(\alpha_{12}, \alpha_{21}\) are competitive effect of the other species; could be the same for both species

Sensitivity analysis

Consider pred-prey BUT what will be the output - if we want to ‘quantify sensitivity’ useful to look at a single value or set of value

For example

• Max Prey Population
• Min Prey Population

Sensitivity Analysis steps

• Generate parameters (LHS, Sobol)
• Metrics function
• Wrapper Function
• Run wrapper function to get metrics for all parameter sets
• Graph and compute sensitivity statistics

NOTE - I’ve also given you an example using Latin Hypercube Sampling * optional review

source("../R/lotvmodK.R")
# lets start with sobol 
library(sensitivity)


# want to learn about sensitivity to growth rate (r) and carrying capacity 
# set the number of parameters
np=200
K = rnorm(mean=150, sd=20, n=np)
rprey = runif(min=0.01, max=0.3, n=np)
alpha= runif(min=0.1, max=0.4, n=np)
eff= rnorm(mean=0.3, sd=0.01, n=np)
pmort= runif(min=0.01, max=0.45, n=np)

X1 = cbind.data.frame(rprey=rprey, K=K, alpha=alpha, eff=eff, pmort=pmort)

# repeat to get our second set of samples
np=200
K = rnorm(mean=150, sd=20, n=np)
rprey = runif(min=0.01, max=0.3, n=np)
alpha= runif(min=0.1, max=0.4, n=np)
eff= rnorm(mean=0.3, sd=0.01, n=np)
pmort= runif(min=0.01, max=0.45, n=np)

X2 = cbind.data.frame(rprey=rprey, K=K, alpha=alpha, eff=eff, pmort=pmort)


# create our sobel object and get sets ofparameters for running the model
sens_PP = sobolSalt(model = NULL,X1, X2, nboot = 300)

# name parameter sets...
colnames(sens_PP$X) = c("rprey","K","alpha","eff","pmort")

# our metrics 
# lets say we  want the maximum and minimum  of both predictor and prey

compute_metrics = function(result) {
  maxprey = max(result$prey)
  maxpred = max(result$pred)
  minprey = min(result$prey)
  minpred = min(result$pred)
return(list(maxprey=maxprey, minprey=minprey, maxpred=maxpred, minpred=minpred))}

# build a wrapper function


p_wrapper = function(rprey,alpha, eff, pmort, K, currpop, days, func) {
    parms = list(rprey=rprey, alpha=alpha, eff=eff, pmort=pmort, K=K)
    result = ode(y=currpop, times=days, func=func, parms=parms) 
    colnames(result)=c("time","prey","pred")
  # get metrics
  metrics=compute_metrics(as.data.frame(result))
  return(metrics)
}


# run our model for all parameters and extract the results
currpop=c(prey=1, pred=1)
days = seq(from=1,to=500)
allresults = as.data.frame(sens_PP$X) %>% pmap(p_wrapper, currpop=currpop, days=days, func=lotvmodK)
 
# take results back to unlisted form
allres = allresults %>% map_dfr(`[`,c("maxprey","minprey","maxpred","minpred"))


# range of response across parameter uncertainty
allresl = allres %>% gather(key="metric",value="pop")
ggplot(allresl, aes(metric, pop))+geom_boxplot()

# dealing with different scales
ggplot(allresl, aes(metric, pop, col=metric))+geom_boxplot()+facet_wrap(~metric, scales="free")

# plot cummulative densities

ggplot(allresl, aes(pop, col=metric))+stat_ecdf(geom="line")+facet_wrap(~metric, scales="free")

# create sobol indices for Max Prey
sens_PP_maxprey = sens_PP %>% sensitivity::tell(y=allres$maxprey)
rownames(sens_PP_maxprey$S)=c("rprey","K","alpha","eff","pmort")
sens_PP_maxprey$S

##         original         bias std. error  min. c.i. max. c.i.
## rprey 0.03183015  0.001526934 0.06091984 -0.0871428 0.1471826
## K     0.02011888  0.001056463 0.06268305 -0.1051848 0.1344842
## alpha 0.40928403 -0.004943837 0.09649863  0.2462706 0.6131552
## eff   0.01657897  0.000982677 0.06247087 -0.1107414 0.1357176
## pmort 0.54727349 -0.001595660 0.04771356  0.4532608 0.6382406

rownames(sens_PP_maxprey$T)=c("rprey","K","alpha","eff","pmort")
sens_PP_maxprey$T

##          original          bias   std. error     min. c.i.   max. c.i.
## rprey 0.012349345  3.243570e-04 0.0028022463  0.0053030919 0.016364154
## K     0.001154020 -2.182323e-05 0.0006471949 -0.0003453104 0.002116712
## alpha 0.467861815 -1.025672e-03 0.0579236062  0.3583067443 0.582321117
## eff   0.003502894  6.578112e-05 0.0006738263  0.0017952034 0.004561289
## pmort 0.667120368 -1.954088e-03 0.0962965284  0.4794595302 0.855046984

Extra Example (LHS)

Here’s an example using Latin HyperCube Sampling

With Partial Rank Correlation Coefficient

source("../R/lotvmodK.R")
library(lhs)
library(epiR)

## Loading required package: survival

## Package epiR 2.0.61 is loaded

## Type help(epi.about) for summary information

## Type browseVignettes(package = 'epiR') to learn how to use epiR for applied epidemiological analyses

## 

# create parameter sets...
factors = c("rprey","K","alpha","eff","pmort")
nsets=500

# create a latin hyper cube
sens_pp = randomLHS(nsets, length(factors))
colnames(sens_pp)=factors

# refine using sampling distributions for each parameter
sens_pp[,"rprey"]= qunif(sens_pp[,"rprey"], min=0.01, max=0.3)
sens_pp[,"K"]= qunif(sens_pp[,"K"], min=10, max=200)
sens_pp[,"alpha"]= qunif(sens_pp[,"alpha"], min=0.1, max=0.4)
sens_pp[,"eff"]= qnorm(sens_pp[,"eff"], mean=0.3, sd=0.01)
sens_pp[,"pmort"]= qunif(sens_pp[,"pmort"], min=0.05, max=0.45)

# lets create a metric and wrapper function as we did for our population models

# first our metrics 
# lets say we  want the maximum and minimum  of both predictor and prey

compute_metrics = function(result) {
  maxprey = max(result$prey)
  maxpred = max(result$pred)
  minprey = min(result$prey)
  minpred = min(result$pred)
return(list(maxprey=maxprey, minprey=minprey, maxpred=maxpred, minpred=minpred))}

# build a wrapper function


p_wrapper = function(rprey,alpha, eff, pmort, K, currpop, days, func) {
    parms = list(rprey=rprey, alpha=alpha, eff=eff, pmort=pmort, K=K)
    result = ode(y=currpop, times=days, func=func, parms=parms) 
    colnames(result)=c("time","prey","pred")
  # get metrics
  metrics=compute_metrics(as.data.frame(result))
  return(metrics)
}


# run our model for all parameters and extract the results
currpop=c(prey=1, pred=1)
days = seq(from=1,to=500)
allresults = as.data.frame(sens_pp) %>% pmap(p_wrapper, currpop=currpop, days=days, func=lotvmodK)
 
# take results back to unlisted form
allres = allresults %>% map_dfr(`[`,c("maxprey","minprey","maxpred","minpred"))


# range of response across parameter uncertainty
allresl = allres %>% gather(key="metric",value="pop")
ggplot(allresl, aes(metric, pop))+geom_boxplot()

# dealing with different scales
ggplot(allresl, aes(metric, pop, col=metric))+geom_boxplot()+facet_wrap(~metric, scales="free")

# plot cummulative densities

ggplot(allresl, aes(pop, col=metric))+stat_ecdf(geom="line")+facet_wrap(~metric, scales="free")

# compute PRCCs using epi.prcc
# lets do first for maxpred

epi.prcc(cbind.data.frame(sens_pp, allres$maxpred))

##     var         est      lower      upper test.statistic  df       p.value
## 1 rprey  0.61052928  0.5408005  0.6802580      17.202812 498  2.134851e-52
## 2     K  0.46890664  0.3911437  0.5466696      11.847276 498  1.066825e-28
## 3 alpha -0.84212897 -0.8896080 -0.7946500     -34.848336 498 1.173191e-135
## 4   eff  0.04578135 -0.0421684  0.1337311       1.022725 498  3.069344e-01
## 5 pmort  0.83426357  0.7857202  0.8828070      33.765881 498 7.387155e-131

# try minprey
epi.prcc(cbind.data.frame(sens_pp, allres$minprey))

##     var         est       lower       upper test.statistic  df       p.value
## 1 rprey  0.83481938  0.78635012  0.88328864     33.8400489 498 3.448977e-131
## 2     K -0.10159453 -0.18918105 -0.01400801     -2.2789664 498  2.309146e-02
## 3 alpha -0.78715593 -0.84145622 -0.73285563    -28.4815381 498 1.356102e-106
## 4   eff  0.04202849 -0.04593577  0.12999276      0.9387337 498  3.483226e-01
## 5 pmort  0.64755379  0.58046401  0.71464358     18.9637535 498  9.116234e-61

And just for fun

• Lorenz Equations (for fluid dynamics), P, R, B are parameters (fixed values), x,y,z variables that change with time that describe how conveciton in the atmosphere works - a cell that is warmed from below and cooled from above

• Developed by Meteorologist Edward Lorenz - early climate model development in 1960s

  • fluid convection
  • x rate of convective overturning
  • y departure from linear horizontal temperature gradient
  • z departure from linear vertical temperature difference

• Lorenz equations are example of dynamic systems that can exhibit stable and chaotic states depending on parameters and initial conditions

Lets look at a Lorenz System with its interesting dynamics

# lorenze
source("../R/lorenz.R")
pars = list(a=10,b=28,c=8/3)
res = ode(func=lorenz, c(x=0.1,y=0,z=0), times=seq(0,50,by=0.01), parms=pars)

ggplot(as.data.frame(res), aes(x,y, col=time))+geom_point()

ggplot(as.data.frame(res), aes(x,z, col=time))+geom_point()

ggplot(as.data.frame(res), aes(y,z, col=time))+geom_point()

resl = as.data.frame(res) %>% gather(key="var", value="value",-time)
ggplot(resl, aes(time, value, col=var))+geom_line()

# try with different initial conditions
pars = list(a=15,b=28,c=8/4)
res = ode(func=lorenz, c(x=0.3,y=5,z=10), times=seq(0,50,by=0.01), parms=pars)

ggplot(as.data.frame(res), aes(x,y, col=time))+geom_point()+scale_colour_gradientn(colours = terrain.colors(10))

ggplot(as.data.frame(res), aes(x,z, col=time))+geom_point()+scale_colour_gradientn(colours = terrain.colors(10))

ggplot(as.data.frame(res), aes(y,z, col=time))+geom_point()+scale_colour_gradientn(colours = terrain.colors(10))

resl = as.data.frame(res) %>% gather(key="var", value="value",-time)
ggplot(resl, aes(time, value, col=var))+geom_line()